CN108572183A - The method for checking equipment and dividing vehicle image - Google Patents
The method for checking equipment and dividing vehicle image Download PDFInfo
- Publication number
- CN108572183A CN108572183A CN201710135023.7A CN201710135023A CN108572183A CN 108572183 A CN108572183 A CN 108572183A CN 201710135023 A CN201710135023 A CN 201710135023A CN 108572183 A CN108572183 A CN 108572183A
- Authority
- CN
- China
- Prior art keywords
- image
- vehicle
- pixel
- convolutional network
- trained
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N23/00—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00
- G01N23/02—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material
- G01N23/06—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material and measuring the absorption
- G01N23/10—Investigating or analysing materials by the use of wave or particle radiation, e.g. X-rays or neutrons, not covered by groups G01N3/00 – G01N17/00, G01N21/00 or G01N22/00 by transmitting the radiation through the material and measuring the absorption the material being confined in a container, e.g. in a luggage X-ray scanners
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/143—Segmentation; Edge detection involving probabilistic approaches, e.g. Markov random field [MRF] modelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10116—X-ray image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30248—Vehicle exterior or interior
Abstract
It discloses a kind of method of dividing vehicle image and checks equipment.X-ray transmission scanning is carried out to vehicle, obtains transmission image;Each pixel of transmission image is carried out using trained convolutional network to add class label;Obtain determining the image of the various pieces of vehicle according to the class label of each pixel.Using above-mentioned scheme, the image of the various pieces of vehicle can be relatively accurately partitioned under the situation that vehicle type is various, situation is complicated.
Description
Technical field
This disclosure relates to radiation monitoring technology, and in particular to a method of checking equipment and dividing vehicle image.
Background technology
As vehicle fleet size also rapidly goes up, it is important as one how efficiently and accurately to carry out the Smart Verify of vehicle
The problem of.It includes the inspection to vehicle various pieces, such as the foreign bodies detection and freight classification of cargo moiety, wheel to examine vehicle
Or the entrainment detection etc. of chassis portion.The basis for carrying out these inspections is that car body understands that is, machine is understanding each of car body
Position is belonging respectively to be partitioned into the different piece of car body on the basis of which part of car body automatically.
The image of field of safety check processing is X-ray transmission figure mostly, and the old car body for car body X-ray transmission figure is managed
Solution method such as parsing etc. is generally based on rule, its effect under situation complexity, the miscellaneous situation of vehicle is poor.
Invention content
In view of one or more problems in the prior art, it is proposed that a kind of inspection equipment and divide vehicle image
The method cut.
According to one aspect of the disclosure, it is proposed that a kind of method of dividing vehicle image, including step:Vehicle is carried out
X-ray transmission scans, and obtains transmission image;List mark is added to each pixel of transmission image using trained convolutional network
Label;Obtain determining the image of the various pieces of vehicle according to the class label of each pixel.
According to some embodiments, convolutional network is trained by following step:The X-ray for obtaining multiple vehicles is saturating
Image is penetrated, as sample image;Pixel-level mark is carried out to the sample image according to the various pieces of vehicle;After mark
Sample image convolutional network is trained.
According to some embodiments, the various pieces of the vehicle include:Wheel, headstock, chassis, compartment.
According to some embodiments, the method further includes at least one of being proceeded as follows to the sample image:
Mirror image, except air part, change of scale, cut into subgraph.
According to some embodiments, the operation for cutting into subgraph includes:Cut the upper left corner, the upper right of the sample image
At least one of angle, the lower left corner, the lower right corner and middle section.
According to some embodiments, the operation for cutting into subgraph includes:The sample image is cut at random.
According to some embodiments, the convolutional network includes at least one of:Full convolutional network, the full convolution of Dilation
Network, Deeplab networks, ConvPP networks.
In another aspect of the present disclosure, it is proposed that a kind of inspection equipment, including:X-ray scanning system, to being examined vehicle
X-ray transmission scanning is carried out, transmission image is obtained;Memory stores the transmission image;Processor is configured to described
It penetrates image and carries out following operation:Each pixel of transmission image is labeled using trained convolutional network;According to each
The mark of a pixel obtains determining the image of the various pieces of vehicle.
According to some embodiments, the processor is configured to be trained convolutional network by following step:It obtains
The X-ray transmission image of multiple vehicles, as sample image;Pixel is carried out to the sample image according to the various pieces of vehicle
Grade mark;Convolutional network is trained using the sample image after mark.
According to some embodiments, the processor is configured to further include proceeding as follows at least the sample image
One of:Mirror image, except air part, change of scale, cut into subgraph.
Using above-mentioned scheme, it can relatively accurately be partitioned into vehicle under the situation that vehicle type is various, situation is complicated
Various pieces image.
Description of the drawings
It for a better understanding of the present invention, will the present invention will be described in detail according to the following drawings:
Fig. 1 shows the structural schematic diagram of the inspection equipment according to the embodiment of the present disclosure;
Fig. 2 is the schematic diagram of the structure for the computing device that description inspection equipment as described in Figure 1 includes;
Fig. 3 is the schematic diagram for the process for describing the training convolutional network according to the embodiment of the present disclosure;
Fig. 4 is the schematic diagram for the process for describing the dividing vehicle image according to the embodiment of the present disclosure;
Fig. 5 shows the schematic diagram of the vehicle transmission image according to the embodiment of the present disclosure;And
Fig. 6 shows the schematic diagram for each part of vehicle partial image that the segmentation according to the embodiment of the present disclosure obtains.
Specific implementation mode
Specific embodiments of the present invention are described more fully below, it should be noted that the embodiments described herein is served only for illustrating
Illustrate, is not intended to restrict the invention.In the following description, in order to provide a thorough understanding of the present invention, a large amount of spies are elaborated
Determine details.It will be apparent, however, to one skilled in the art that:This hair need not be carried out using these specific details
It is bright.In other instances, in order to avoid obscuring the present invention, well known structure, material or method are not specifically described.
Throughout the specification, meaning is referred to " one embodiment ", " embodiment ", " example " or " example "
It:A particular feature, structure, or characteristic described in conjunction with this embodiment or example is comprised at least one embodiment of the present invention.
Therefore, the phrase " in one embodiment ", " in embodiment ", " example " occurred in each place of the whole instruction
Or " example " is not necessarily all referring to the same embodiment or example.Furthermore, it is possible to will be specific with any combination appropriate and/or sub-portfolio
Feature, structure or characteristic combine in one or more embodiments or example.In addition, those of ordinary skill in the art should manage
Solution, term "and/or" used herein includes any and all combinations for the project that one or more correlations are listed.
In view of such as rule-based understanding method to vehicle X-ray transmission figure existing in the prior art in vehicle type
The poor deficiency of effect in the case of numerous carries out car body understanding, in vehicle using deep learning method to vehicle X-ray transmission figure
X-ray transmission figure on be partitioned into the different piece of car body, such as wheel, headstock, chassis, container profile etc..Specifically,
X-ray transmission scanning is carried out to the vehicle of such as container truck etc, obtains transmission image.Then, utilization is trained
Such as the convolutional network of full convolutional network (FCN) adds class label to each pixel of transmission image.Next, according to each picture
The class label of element determines the image of the various pieces of vehicle.Apply to deep learning, phase by the above method, such as by GPU
Need manual extraction feature compared with traditional machine learning method, deep learning by the feature that learns to obtain come intelligent classification,
Detection, semantic segmentation are more and more accurate.
Specifically, the existing car body understanding method for vehicle X-ray transmission figure is rule-based, and manpower is very
Difficult limit strictly all rules.In comparison, feature extraction and pixel classifications are merged by end-to-end deep learning.With big
Under conditions of scale sample training, allow the feature that machine oneself learns that there is universality, to which generalization ability is stronger.Thus in vehicle
Effect is more preferable under the situation that type type is various, situation is complicated.It in a particular application, can be by the image for the different piece being partitioned into
For subsequent intelligent measurement.The intelligent measurement that can be helped to realize vehicle to a certain extent using above-mentioned scheme, is helped
In efficiently and safely checking vehicle.
Fig. 1 shows the structural schematic diagram of the inspection equipment according to the embodiment of the present disclosure.As shown in Figure 1, according to the disclosure
The inspection equipment 100 of embodiment includes x-ray source 110, detector 130, data acquisition device 150, controller 140 and calculates
Equipment 160 carries out safety inspection to the inspected object 120 of such as container truck etc, such as judges wherein whether include
Dangerous material/or suspicious object.Although in this embodiment, that detector 130 and data acquisition device 150 is described separately,
It is that it should be appreciated by those skilled in the art they can also be integrated to referred to as X-ray detection and data acquisition equipment.
According to some embodiments, above-mentioned x-ray source 110 can be isotope, can also be X-ray machine or accelerator etc..X
Radiographic source 110 can be single energy, can also be dual intensity.In this way, passing through x-ray source 110 and detector 150 and controller 140
Transmission scan is carried out to inspected object 120 with computing device 160, obtains detection data.Such as it advances in inspected object 120
In the process, operating personnel send out instruction, order X is penetrated by means of the human-computer interaction interface of computing device 160 by controller 140
Line source 110 sends out ray, is received by detector 130 and data acquisition equipment 150 after inspected object 120, and pass through
Computing device 160 handles data, on the one hand can obtain transmission image, on the other hand can be by means of trained convolution
Network is split the transmission image of vehicle, obtains the image of the various pieces of vehicle, to facilitate subsequent dangerous material/can
Doubt the inspection of article.In this way, can determine that suspicious object exists after the characteristic value for example by comparing such as atomic number etc
In container or in which position of car body, inspection personnel is facilitated to find out the position of dangerous material/suspicious object.In addition, also may be used
With based on the car body for containing entrainment image and not between the image of each section of the car body comprising entrainment it is existing
Difference further judges whether some part in car body has entrainment.
Fig. 2 shows the structural schematic diagrams of computing device as shown in Figure 1.As shown in Fig. 2, the letter that detector 130 detects
It number is acquired by data collector, data are stored in by interface unit 167 and bus 163 in memory 161.Read-only memory
(ROM) configuration information and program of computer data processor are stored in 162.Random access memory (RAM) 163 is used for
Various data are kept in 165 course of work of processor.In addition, being also stored in memory 161 for carrying out data processing
Computer program, such as artificial neural network program, Object Classification program and image processing program etc..Internal bus 163 connects
Connect above-mentioned memory 161, read-only memory 162, random access memory 163, input unit 164, processor 165, display
Device 166 and interface unit 167.
After the operational order that user is inputted by the input unit 164 of such as keyboard and mouse etc, computer program
Instruction code command processor 165 execute scheduled data processing algorithm, after obtaining data processed result, shown
In the display device 167 of such as LCD display etc, or output is handled directly in the form of such as printing etc hard copy
As a result.
Before being trained to full convolutional network, needs to establish the database for training, then be rolled up entirely using sample training
Product network.For example, in embodiment of the disclosure, manually carrying out Pixel-level mark to a large amount of known car body transmission images first
Note, then carries out extensive sample end-to-end training with the car body understanding method based on deep learning.
Specifically, including to a large amount of car body transmission images progress Pixel-level mark:(1) determine that car body transmission image includes
Those classifications;(2) great amount of images is extracted from car body transmission image;(3) use manpower to each of the transmission image that extracts
Its upper generic of pixel mark, to obtain and former transmission image label figure of the same size;(4) the label figure that will have been marked
It is preserved with former transmission image, is used as training sample.
It, can be with the car body understanding method based on deep learning to extensive sample when being trained to full convolutional network
Carry out end-to-end training, used in deep learning method be based on full convolutional network (FCN), on its basis
Also some improved network structures.
Specifically, FCN is a kind of convolutional neural networks (CNN) model being used for doing semantic segmentation, it is by that will divide
The full articulamentum of class CNN networks is re-interpreted as network structure obtained from convolutional layer, this so that the input picture of network can be with
It is arbitrary size, and prediction result is reduced to original image size finally by up-sampling.
In addition, on the basis of FCN, a kind of improved FCN, i.e. Dilation FCN can also be used.It is by FCN most
Latter two Pooling layers of step-length becomes 1 by 2, and the receptive field of output layer is then made by way of adding hole to convolution kernel not
Become, further, it is multiple dimensioned to achieve the purpose that different rate is set.
In addition, on the basis of Dilation FCN, it can also be improved:Image is passed through trained
Dilation FCN networks obtain the class probability distribution of each pixel, in conjunction with the information of image itself, with a condition with
Airport (CRF) is distributed to optimize the class probability of each pixel in image, finally obtains the higher prediction label figure of resolution ratio.
Dilation FCN are referred to as DeepLab plus the method for CRF.CRF in DeepLab itself can be construed to a cycle
Neural network (RNN).CRF is construed to RNN to be also an advantage that, that is, certain weight parameters in CRF can pass through RNN
Study obtains.
According to some embodiments, a kind of network C onvPP can also be used, it trains a kind of special using label itself
Convolutional network, using this special convolutional network train come feature, train the feature come in conjunction with FCN before upper and come pair
Each pixel, which is done, classifies, and the result obtained in this way will be better than FCN.
In some embodiments of the present disclosure, the training of great amount of samples is carried out to above-mentioned convolutional network.In other embodiment
In, it can also reach much the same with above method as a result, such as Low Shot networks using a small amount of sample.
Fig. 3 is the schematic diagram for the process for describing the training convolutional network according to the embodiment of the present disclosure.As shown in figure 3, in step
Rapid S310 obtains the transmission image of the vehicle of such as container truck etc, as sample image.Then, in step S320, lead to
Cross the mark that Pixel-level is manually carried out to sample image.For example, determining that vehicle transmission image includes those classifications, then from car body
It is manually marked, is preserved as sample after extracting great amount of images in image.
Specifically, according to actual conditions, vehicle transmission image can be divided into 6 classes:Air, cargo, container, wheel
Exterior feature, chassis, headstock, wheel.Air should be divided into air part outside air in a car compartment part and compartment originally, but rise for simplicity
See, it is unified that they are all designated as air.
Then, different automobile types are randomly selected with the image of equivalent, the type of sample is more, quantity is bigger, the net trained
Network generalization ability is stronger, and less training sample can lead to over-fitting.
Next, with manpower to its upper generic of each pixel mark of the transmission image extracted, to obtain
With former transmission image label figure of the same size.If original image height and width are respectively H and W, then the label figure marked height and
Wide is also respectively H and W, wherein the value range of each pixel is { 0,1,2,3,4,5 }, respectively represents the pixel in artwork and corresponds to
Pixel belong to which kind of in 6 classes.
Finally, the label figure marked and former transmission image are preserved, is used as training sample.In some embodiments
In, sample can be increased by operating as follows quantity/quickening arithmetic speed.
Mirror image operation:For enlarged sample amount, mirror image can be done to the image in sample, this can make sample size expand
One times.
Except air-operated:In order to improve efficiency, the air portion cutting of image surrounding is removed using label figure, is obtained opposite
Thus smaller image can accelerate training speed when training.
Change of scale:If computing resource is insufficient or in order to improve efficiency, sample image can be contracted by some proportion
It is small, such as narrow down to original 1/5.
Training sample amount can also be expanded by cutting picture:
Furthermore, it is possible to the upper left corner, the upper right corner, the lower left corner, the lower right corner and the middle section of one figure of cutting are selected, wherein
Each section respectively accounts for 70% size of artwork.In addition artwork, at this moment can expand 5 times again by sample size.Alternatively, can also be random
Cutting, then can scheme upper random cutting n times at every, adds artwork at this moment if it is desired to sample size is expanded N times.
According to the other embodiment of the disclosure, this step of cutting image can also be when training network at data layers
It does, can be saved down in this way for storing the hard drive space for the picture being cut into.
In step S330, the sample image training convolutional network after mark is utilized.For example, with the car body based on deep learning
Understanding method carries out extensive sample end-to-end training.
● the training based on FCN
FCN is obtained by the full articulamentum of sorter network is re-interpreted as convolutional layer.Again after solution, it allows to input
Image is arbitrary size.By taking VGG networks as an example, after any one image is by FCN editions VGG networks, length and width are reduced into artwork
1/32, then again plus one layer of warp lamination, up-sample as artwork size.Since by artwork, down-sampled 32 times up-sample again as original
Figure size can lose many detailed information of artwork, therefore pool4 and pool5 layers of feature can be also used in classification, with
Improve the resolution ratio of image.
● the training based on DeepLab
It is trained Dilation FCN first.The optional network structures of Dilation FCN have several:V1、LargeFOV、
MSc-LargeFOV、ASPP、Res101。
The basic network of V1, LargeFOV, MSc-LargeFOV, ASPP are VGG.It, will for the VGG networks of FCN versions
Pool4 and pool5 layers of stride becomes 1 from 2, and the convolution kernel rate of the convolutional layer between pool4 and pool5 layers is set as
2.On this basis, the convolution kernel rate of the convolutional layer after pool5 layers is set as 4 by V1, to make output layer maintain and original
The same receptive fields of FCN of beginning.LargeFOV contributes to the object of identification scale bigger in view of the receptive field of bigger, thus will
The convolution kernel rate of convolutional layer after pool5 layers sets to obtain bigger, such as 12.MSc-LargeFOV is on the bases LargeFOV
On, allow the feature of bottom to also assist in last pixel classifications.ASPP considers multiple dimensioned, and 4 have been divided after pool5 layers,
The convolution kernel of convolutional layer takes different rate, such as first rate=6, second rate=12, third branch in each branch
Rate=18, the 4th rate=24, this makes each pixel of last output layer there are four the receptive field of scale.
Res101 is obtained by changing 101 layers of classification residual error net (resnet).Amending method is similar with VGG.
It is trained with deep learning frame caffe.In two stages:First stage, each iteration, for the figure of input
Picture therefrom fixes the subgraph of length and width for training with machine-cut;Second stage trains the parameter base come in first stage
It is finely tuned on plinth, each iteration inputs an image for training.
Then, training CRF.Although the effect of Dilation FCN is better than original FCN, due to output layer reality
On down-sampled 8 times obtain, so 8 times of up-sampling is obtaining and the equirotal label figure of original image resolution ratio compared with
It is low.In order to improve resolution ratio, CRF can be added after Dilation FCN, there are two types of methods, and CRF is added:(1) by CRF independences
In training sample, it is not involved in training, with the parameter of cross validation obtained in CRF.In prognostic chart picture, image is first passed through into instruction
The Dilation FCN networks perfected, then obtained class probability is distributed to the information for combining upper original image, iteration obtains more
Accurately class probability is distributed.(2) CRF participates in training:CRF is construed to RNN, is connected to behind Dilation FCN, to
Can be to its end-to-end training, the parameter in wherein CRF can be obtained by training.
Fig. 4 is the schematic diagram for the process for describing the dividing vehicle image according to the embodiment of the present disclosure.As shown in figure 4, in step
Rapid S410 carries out X-ray scanning to being examined vehicle by system shown in FIG. 1, obtains transmission image.Herein, also suspicious
Transmission image is carried out to remove air-treatment, noise reduction, the operations such as normalization.
In step S420, class label is added to the pixel of transmission image using the convolutional network mark of above-mentioned training.For example, will
Transmission image is input in trained convolutional network, obtains the class label of each pixel in image, such as the difference to car body
Partial pixel does corresponding label, for example, wheel is 1, headstock 2, chassis 3, compartment 4.
In step S430, the image of vehicle various pieces is determined according to the class label of vehicle pixel.Such as to different marks
The pixel of label assigns different colors, or the car body of different piece is fenced up with closed curve, indicates the vehicle identified
Body portion, to facilitate the inspection of subsequent dangerous material/suspicious object.In this way, ought for example by compare such as atomic number it
It can determine that suspicious object in container or in which position of car body, facilitates inspection personnel to find out danger after the characteristic value of class
The position of dangerous product/suspicious object.In addition it is also possible to image based on the car body for containing entrainment with not comprising entrainment
Existing difference further judges whether some part in car body has entrainment between the image of each section of car body.
Fig. 5 shows the schematic diagram of the vehicle transmission image according to the embodiment of the present disclosure.Fig. 6 is shown according to disclosure reality
Apply the schematic diagram for each part of vehicle partial image that the segmentation of example obtains.As shown in fig. 6, passing through the reason to being examined vehicle transmission image
Solution, by car body image segmentation at parts such as wheel, headstock, chassis, compartments.In addition, also suspicious give different body portions as needed
The pixel divided assigns different colors, or the car body of different piece is fenced up with closed curve, indicates the vehicle identified
Body portion.
According to above-described embodiment, under conditions of with extensive sample training, the feature that machine oneself learns is allowed to have general
Adaptive, to which generalization ability is stronger.Thus effect is more preferable under the situation that vehicle type is various, situation is complicated.In concrete application
In, the image for the different piece being partitioned into can be used for subsequent intelligent measurement.It can be in certain journey using above-mentioned scheme
The intelligent measurement of vehicle is realized in help on degree, helps efficiently and safely to check vehicle.
Above detailed description has elaborated to check equipment and image by using schematic diagram, flow chart and/or example
Numerous embodiments of dividing method.Include one or more functions and/or operation in this schematic diagram, flow chart and/or example
In the case of, it will be understood by those skilled in the art that each function and/or operation in this schematic diagram, flow chart or example can
Individually and/or jointly to be realized by various structures, hardware, software, firmware or they substantial arbitrary combination.At one
In embodiment, if the stem portion of theme described in the embodiment of the present invention can be compiled by application-specific integrated circuit (ASIC), scene
Journey gate array (FPGA), digital signal processor (DSP) or other integrated formats are realized.However, those skilled in the art answer
It recognizes, some aspects of embodiments disclosed herein can equally be realized in integrated circuit on the whole or partly
In, the one or more computer programs for being embodied as running on one or more computer are (for example, be embodied as at one or more
The one or more programs run in platform computer system), it is embodied as run on the one or more processors one or more
A program (for example, being embodied as the one or more programs run in one or more microprocessors), is embodied as firmware, or
Substantially be embodied as the arbitrary combination of aforesaid way, and those skilled in the art are according to the disclosure, will be provided with design circuit and/
Or the ability of write-in software and/or firmware code.In addition, it would be recognized by those skilled in the art that the machine of theme described in the disclosure
System can be distributed as the program product of diversified forms, and no matter actually be used for executing the signal bearing medium of distribution
How is concrete type, and the exemplary embodiment of theme described in the disclosure is applicable in.The example of signal bearing medium includes but unlimited
In:Recordable-type media, such as floppy disk, hard disk drive, compact-disc (CD), digital versatile disc (DVD), digital magnetic tape, computer
Memory etc.;And transmission type media, such as number and/or analogue communication medium are (for example, optical fiber cable, waveguide, wire communication chain
Road, wireless communication link etc.).
Although exemplary embodiment describing the present invention with reference to several, it is to be understood that, term used is explanation and shows
Example property, term and not restrictive.The spirit or reality that can be embodied in a variety of forms without departing from invention due to the present invention
Matter, it should therefore be appreciated that above-described embodiment is not limited to any details above-mentioned, and should be spiritual defined by appended claims
Accompanying is all should be with the whole variations and remodeling widely explained, therefore fallen into claim or its equivalent scope in range to weigh
Profit requires to be covered.
Claims (11)
1. a kind of method of dividing vehicle image, including step:
X-ray transmission scanning is carried out to vehicle, obtains transmission image;
Class label is added to each pixel of transmission image using trained convolutional network;And
The image of the various pieces of vehicle is determined according to the class label of each pixel.
2. the method as described in claim 1, wherein being trained to convolutional network by following step:
The X-ray transmission image for obtaining multiple vehicles, as sample image;
Pixel-level mark is carried out to the sample image according to the various pieces of vehicle;
Convolutional network is trained using the sample image after mark.
3. the method as described in claim 1, wherein the various pieces of the vehicle include:Wheel, headstock, chassis, compartment.
4. method as claimed in claim 2, wherein further include at least one of being proceeded as follows to the sample image:
Mirror image, except air part, change of scale, cut into subgraph.
5. method as claimed in claim 4, wherein the operation for cutting into subgraph includes:
Cut at least one of the upper left corner, the upper right corner, the lower left corner, the lower right corner and the middle section of the sample image.
6. method as claimed in claim 4, wherein the operation for cutting into subgraph includes:
The sample image is cut at random.
7. the method as described in claim 1, wherein the convolutional network includes at least one of:
Full convolutional network, the full convolutional networks of Dilation, Deeplab networks, ConvPP networks.
8. a kind of inspection equipment, including:
X-ray scanning system carries out X-ray transmission scanning to being examined vehicle, obtains transmission image;
Memory stores the transmission image;
Processor is configured to carry out following operation to the transmission image:
Class label is added to each pixel of transmission image using trained convolutional network;
The image of the various pieces of vehicle is determined according to the class label of each pixel.
9. equipment is checked as described in claim 1, wherein the processor is configured to through following step to convolutional network
It is trained:
The X-ray transmission image for obtaining multiple vehicles, as sample image;
Pixel-level mark is carried out to the sample image according to the various pieces of vehicle;
Convolutional network is trained using the sample image after mark.
10. equipment is checked as described in claim 1, wherein the various pieces of the vehicle include:
Wheel, headstock, chassis, compartment.
11. as claimed in claim 9 check equipment, wherein the processor be configured to further include to the sample image into
At least one of following operation of row:
Mirror image, except air part, change of scale, cut into subgraph.
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710135023.7A CN108572183B (en) | 2017-03-08 | 2017-03-08 | Inspection apparatus and method of segmenting vehicle image |
US15/913,222 US10796436B2 (en) | 2017-03-08 | 2018-03-06 | Inspection apparatuses and methods for segmenting an image of a vehicle |
EP18160385.3A EP3373246A1 (en) | 2017-03-08 | 2018-03-07 | Inspection apparatuses and methods for segmenting an image of a vehicle |
JP2018041778A JP6757758B2 (en) | 2017-03-08 | 2018-03-08 | How to split images of inspection equipment and vehicles |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710135023.7A CN108572183B (en) | 2017-03-08 | 2017-03-08 | Inspection apparatus and method of segmenting vehicle image |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108572183A true CN108572183A (en) | 2018-09-25 |
CN108572183B CN108572183B (en) | 2021-11-30 |
Family
ID=61598899
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710135023.7A Active CN108572183B (en) | 2017-03-08 | 2017-03-08 | Inspection apparatus and method of segmenting vehicle image |
Country Status (4)
Country | Link |
---|---|
US (1) | US10796436B2 (en) |
EP (1) | EP3373246A1 (en) |
JP (1) | JP6757758B2 (en) |
CN (1) | CN108572183B (en) |
Families Citing this family (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109360210B (en) | 2018-10-16 | 2019-10-25 | 腾讯科技(深圳)有限公司 | Image partition method, device, computer equipment and storage medium |
CN111161274B (en) * | 2018-11-08 | 2023-07-07 | 上海市第六人民医院 | Abdominal image segmentation method and computer equipment |
JP7320972B2 (en) * | 2019-04-08 | 2023-08-04 | 株式会社日立ハイテク | Image processing device, automatic analysis system and image processing method |
CN110309770B (en) * | 2019-06-28 | 2022-09-30 | 华侨大学 | Vehicle re-identification method based on four-tuple loss metric learning |
JP7250331B2 (en) * | 2019-07-05 | 2023-04-03 | 株式会社イシダ | Image generation device, inspection device and learning device |
CN113850275A (en) * | 2019-09-27 | 2021-12-28 | 深圳市商汤科技有限公司 | Image processing method, image processing device, electronic equipment and storage medium |
CN110956081B (en) * | 2019-10-14 | 2023-05-23 | 广东星舆科技有限公司 | Method and device for identifying position relationship between vehicle and traffic marking and storage medium |
CN111767831B (en) * | 2020-06-28 | 2024-01-12 | 北京百度网讯科技有限公司 | Method, apparatus, device and storage medium for processing image |
CN111860361B (en) * | 2020-07-24 | 2023-07-25 | 吉林大学 | Automatic identifier and identification method for green channel cargo scanning image entrainment |
CN111968028A (en) * | 2020-08-14 | 2020-11-20 | 北京字节跳动网络技术有限公司 | Image generation method, device, equipment and computer readable medium |
CN112614097B (en) * | 2020-12-16 | 2022-02-01 | 哈尔滨市科佳通用机电股份有限公司 | Method for detecting foreign matter on axle box rotating arm of railway train |
Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1781459A (en) * | 2004-12-01 | 2006-06-07 | Ge医疗系统环球技术有限公司 | Dose evaluating method and X-ray CT apparatus |
US20080152082A1 (en) * | 2006-08-16 | 2008-06-26 | Michel Bouchard | Method and apparatus for use in security screening providing incremental display of threat detection information and security system incorporating same |
CN101339652A (en) * | 2007-12-28 | 2009-01-07 | 中国人民解放军海军航空工程学院 | Solid engines CT image division method |
US20090175411A1 (en) * | 2006-07-20 | 2009-07-09 | Dan Gudmundson | Methods and systems for use in security screening, with parallel processing capability |
CN101627918A (en) * | 2008-07-18 | 2010-01-20 | Ge医疗系统环球技术有限公司 | Method and device for compressing CT images |
CN101878435A (en) * | 2007-09-28 | 2010-11-03 | 莫弗探测公司 | Systems and methods for reducing false alarms in detection systems |
CN102540445A (en) * | 2010-12-06 | 2012-07-04 | 索尼公司 | Microscope, region determining method, and program |
CN102539460A (en) * | 2012-01-06 | 2012-07-04 | 公安部第一研究所 | Projection center-of-rotation positioning method of computerized tomography (CT) system |
CN102590239A (en) * | 2007-10-05 | 2012-07-18 | 清华大学 | Method and equipment for determining computerized tomography (CT) scanning position in drug inspection system |
CN103886318A (en) * | 2014-03-31 | 2014-06-25 | 武汉天仁影像科技有限公司 | Method for extracting and analyzing nidus areas in pneumoconiosis gross imaging |
CN104616032A (en) * | 2015-01-30 | 2015-05-13 | 浙江工商大学 | Multi-camera system target matching method based on deep-convolution neural network |
CN104751163A (en) * | 2013-12-27 | 2015-07-01 | 同方威视技术股份有限公司 | Fluoroscopy examination system and method for carrying out automatic classification recognition on goods |
CN106023220A (en) * | 2016-05-26 | 2016-10-12 | 史方 | Vehicle exterior part image segmentation method based on deep learning |
Family Cites Families (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2525228B (en) * | 2014-04-16 | 2020-05-06 | Smiths Heimann Sas | Identification or determination of a load based on texture |
JP6440303B2 (en) * | 2014-12-02 | 2018-12-19 | エヌ・ティ・ティ・コムウェア株式会社 | Object recognition device, object recognition method, and program |
CN105808555B (en) * | 2014-12-30 | 2019-07-26 | 清华大学 | Check the method and system of cargo |
JP6546271B2 (en) * | 2015-04-02 | 2019-07-17 | 株式会社日立製作所 | Image processing apparatus, object detection apparatus, and image processing method |
US9514391B2 (en) * | 2015-04-20 | 2016-12-06 | Xerox Corporation | Fisher vectors meet neural networks: a hybrid visual classification architecture |
JP2017004350A (en) | 2015-06-12 | 2017-01-05 | 株式会社リコー | Image processing system, image processing method and program |
-
2017
- 2017-03-08 CN CN201710135023.7A patent/CN108572183B/en active Active
-
2018
- 2018-03-06 US US15/913,222 patent/US10796436B2/en active Active
- 2018-03-07 EP EP18160385.3A patent/EP3373246A1/en not_active Withdrawn
- 2018-03-08 JP JP2018041778A patent/JP6757758B2/en active Active
Patent Citations (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN1781459A (en) * | 2004-12-01 | 2006-06-07 | Ge医疗系统环球技术有限公司 | Dose evaluating method and X-ray CT apparatus |
US20090175411A1 (en) * | 2006-07-20 | 2009-07-09 | Dan Gudmundson | Methods and systems for use in security screening, with parallel processing capability |
US20080152082A1 (en) * | 2006-08-16 | 2008-06-26 | Michel Bouchard | Method and apparatus for use in security screening providing incremental display of threat detection information and security system incorporating same |
CN101878435A (en) * | 2007-09-28 | 2010-11-03 | 莫弗探测公司 | Systems and methods for reducing false alarms in detection systems |
CN102590239A (en) * | 2007-10-05 | 2012-07-18 | 清华大学 | Method and equipment for determining computerized tomography (CT) scanning position in drug inspection system |
CN101339652A (en) * | 2007-12-28 | 2009-01-07 | 中国人民解放军海军航空工程学院 | Solid engines CT image division method |
CN101627918A (en) * | 2008-07-18 | 2010-01-20 | Ge医疗系统环球技术有限公司 | Method and device for compressing CT images |
CN102540445A (en) * | 2010-12-06 | 2012-07-04 | 索尼公司 | Microscope, region determining method, and program |
CN102539460A (en) * | 2012-01-06 | 2012-07-04 | 公安部第一研究所 | Projection center-of-rotation positioning method of computerized tomography (CT) system |
CN104751163A (en) * | 2013-12-27 | 2015-07-01 | 同方威视技术股份有限公司 | Fluoroscopy examination system and method for carrying out automatic classification recognition on goods |
CN103886318A (en) * | 2014-03-31 | 2014-06-25 | 武汉天仁影像科技有限公司 | Method for extracting and analyzing nidus areas in pneumoconiosis gross imaging |
CN104616032A (en) * | 2015-01-30 | 2015-05-13 | 浙江工商大学 | Multi-camera system target matching method based on deep-convolution neural network |
CN106023220A (en) * | 2016-05-26 | 2016-10-12 | 史方 | Vehicle exterior part image segmentation method based on deep learning |
Non-Patent Citations (4)
Title |
---|
JONATHAN LONG 等: "Fully Convolutional Networks for Semantic Segmentation", 《PROCEEDINGS OF THE 2015 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION》 * |
NICOLAS JACCARD 等: "Tackling the X-ray cargo inspection challenge using machine learning", 《PROC. SPIE》 * |
VICTOR J.ORPHAN 等: "Advanced g ray technology for scanning cargo containers", 《APPLIED RADIATION AND ISOTOPES》 * |
WENHAO LU 等: "Parsing Semantic Parts of Cars Using Graphical Models and Segment Appearance Consistency", 《ARXIV》 * |
Also Published As
Publication number | Publication date |
---|---|
US20180260959A1 (en) | 2018-09-13 |
JP2018146587A (en) | 2018-09-20 |
US10796436B2 (en) | 2020-10-06 |
EP3373246A1 (en) | 2018-09-12 |
JP6757758B2 (en) | 2020-09-23 |
CN108572183B (en) | 2021-11-30 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108572183A (en) | The method for checking equipment and dividing vehicle image | |
Maeda et al. | Generative adversarial network for road damage detection | |
US20190050981A1 (en) | System and method for automatically detecting a target object from a 3d image | |
CN112330591B (en) | Steel rail surface defect detection method and device capable of achieving sample-less learning | |
US10074166B2 (en) | Systems and methods for inspecting cargoes | |
CN109583562A (en) | SGCNN: the convolutional neural networks based on figure of structure | |
CN109117814B (en) | Image processing method, image processing apparatus, electronic device, and medium | |
US11748975B2 (en) | Method and device for optimizing object-class model based on neural network | |
Gorokhovatskyi et al. | Explanation of CNN image classifiers with hiding parts | |
CN115034200A (en) | Drawing information extraction method and device, electronic equipment and storage medium | |
CN114998220A (en) | Tongue image detection and positioning method based on improved Tiny-YOLO v4 natural environment | |
CN116645586A (en) | Port container damage detection method and system based on improved YOLOv5 | |
CN115019133A (en) | Method and system for detecting weak target in image based on self-training and label anti-noise | |
CN113936220A (en) | Image processing method, storage medium, electronic device, and image processing apparatus | |
CN113095404A (en) | X-ray contraband detection method based on front and back background partial convolution neural network | |
Siriborvornratanakul | A deep learning based road distress visual inspection system using Modified U-Net | |
Hett et al. | Graph of hippocampal subfields grading for Alzheimer’s disease prediction | |
Alves et al. | Image segmentation based on ultimate levelings: From attribute filters to machine learning strategies | |
Artan et al. | Car damage analysis for insurance market using convolutional neural networks | |
CN111461152A (en) | Cargo detection method and device, electronic equipment and computer readable medium | |
CN113343767A (en) | Logistics illegal operation identification method, device, equipment and storage medium | |
CN112036312A (en) | Quality inspection scene recognition method and device, computer equipment and readable storage medium | |
CN110310897A (en) | The guidance inspection of semiconductor wafer based on space density analysis | |
Ferreira et al. | Conformity Assessment of Informative Labels in Car Engine Compartment with Deep Learning Models | |
Xu | A fusion-based approach to deep-learning and edge-cutting algorithms for identification and color recognition of traffic lights |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |